On Friday at Meta's @Scale conference, Boris Cherny, creator of Claude Code, revealed that the next big leap in artificial intelligence is not just automating coding, but looping it endlessly. Agents are now prompting other agents to continuously refine code architecture and eliminate redundancies.

Cherny explained that while transitioning from manual coding to agent-written code was a significant step, loops represent another massive advancement: 'As big as the step from source code to agents was, loops are just as important and as big a step.' He highlighted two key loops in his work—one improving code architecture, the other identifying duplicated abstractions for unification. These loops submit pull requests like any human coder, ensuring constant improvement.

The idea may seem daunting, but it's not entirely new. Recursive loops are common in computer science education and have been a part of AI since its inception. One popular loop is the Ralph Loop, which constantly checks if the model has achieved its goal, preventing tasks from running indefinitely. This approach ensures that models don't get stuck, making them more efficient.

Another aspect of this development is the push for more test-time compute. As Noam Brown from OpenAI observed, modern models can solve almost any problem with sufficient computation. Loops allow AI to keep improving incrementally until it reaches a desired threshold or continues indefinitely if there's enough computational power available.

The downside is that these loops are resource-intensive and could be costly for some users. Anthropic, which sells tokens, might find this model profitable, but for others, the expenses could make it a less attractive option. However, with proper setup and oversight, the benefits of continuous improvement might outweigh the costs, making loops a promising development in AI.